Speech synthesis (TTS: Text-to-Speech) generates artificial speech from text input. In telephony, it powers IVR voice guidance, spoken playback of voicemail transcripts, and screen readers for people with visual impairments.
Deep learning changed how natural synthetic speech can sound. Early synthesisers sounded mechanical, but WaveNet, published in 2016, generated the audio waveform itself one sample at a time, and VALL-E, published in 2023, reproduced a speaker's voice from a recording of roughly three seconds. Emotional delivery, intonation control, and imitation of a particular voice are all within reach, and the technology is used for narration and automated telephone responses (as of August 2026).
Misuse is a real concern. Voice clone scams build a convincing imitation of someone's voice from a few seconds of audio and use it to impersonate a family member asking for money. What sits at the core of voice cloning is speech synthesis - text read aloud in a voice the model has learned. Voice recognition, which turns speech into text, is a separate line of technology; pairing the two can produce an automated responder that transcribes what the caller says and answers in a synthetic voice, but in the scam context recognition matters mainly on the authentication side, where a voice is used to confirm identity (voiceprint authentication).
Defences include agreeing on a family code word, hanging up and calling back yourself whenever money comes up, and pairing voice authentication with multi-factor methods such as one-time passwords. See AI voice clone scams for how the tactics work in detail.